AI Foundry TPM Limits Prevent Stable, Mission-Critical Use

Question

AI Foundry TPM Limits Prevent Stable, Mission-Critical Use

Schönwald, Alexander 120

Hello everyone,

I have already raised concerns several times in this forum about how Microsoft handles the rollout and provisioning of models and the associated token limits in AI Foundry. The number of TPM provided is so limited that even a single developer can quickly hit the ceiling. Requests for additional TPM often take an unreasonably long time — if you receive a response at all. And by the time increased limits are finally granted, a newer version of the model is already available in Foundry, forcing the entire process to start over again.

Today’s outage of the Sweden region in Foundry highlights this issue once more. Due to the restricted TPM limits, we are forced to rely on a single region and we couldn't reroute the requests to other regions, forcing us to shut down our services until Sweden is fixed. It is simply not feasible to wait for approval requests every time we want to use another region. As a result, it becomes obvious that building stable and reliable services — especially for mission-critical production workloads — is practically impossible under these conditions.

Within our company, we rely heavily on AI Foundry as a core part of our AI strategy. However, the current approach to token allocation clearly needs to be reworked for professional and enterprise use cases. I fully understand limiting access for private individuals, but why are the same restrictions applied to paying companies and long-standing customers?

Amit Tyagi 135 Reputation points

2026-01-27T17:25:09.89+00:00

Hi Alexander,

You’re raising a very real concern that many teams hit when moving from PoC to production with Azure AI Foundry. Two key pain areas stand out:

1) TPM limits are capacity-driven, not billing-driven

Even for paid subscriptions, TPM is allocated based on regional GPU capacity and model demand, which is why limits are model and region-specific and don’t automatically carry forward to newer model versions. This works for experimentation, but becomes restrictive for mission-critical production workloads. Ref: Docs: https://learn.microsoft.com/azure/ai-services/openai/quotas-limits

2) TPM constraints block standard multi-region resiliency

Because secondary regions often lack pre-approved TPM, teams can be forced into single-region dependency, making it difficult to follow Azure’s own HA/DR best practices. During regional outages, this can lead to full service disruption even when customers are willing to design multi-region architectures. Ref: Docs: https://learn.microsoft.com/azure/architecture/framework/resiliency/overview

Bottom line: AI Foundry can run production workloads today, but only with careful capacity planning and early multi-region approvals. The current TPM and approval flow is not yet optimized for always-on, enterprise-grade resiliency — and this is exactly where platform evolution is needed.

Thanks for clearly articulating this — it’s valuable feedback for the service.
Amit Tyagi 135 Reputation points

2026-01-27T17:27:11.2933333+00:00
This issue is not a request for unlimited capacity, but for predictable, enterprise-grade operability.

Specifically, the current TPM and region-scoped approval model in Azure AI Foundry makes it difficult to design resilient, mission-critical systems, even for paying enterprise customers:

TPM approvals are model- and region-specific, requiring repeated requests with every new model version.

Limited baseline TPM prevents pre-provisioning secondary regions, blocking standard HA/DR patterns.

Regional outages (such as the recent Sweden incident) can lead to complete service disruption, even when customers are willing to operate multi-region.

This effectively places production workloads in a single-region dependency, which is inconsistent with Azure’s general reliability and resiliency guidance.

Request for consideration:

Subscription- or tenant-level TPM allocation models for enterprise customers

Faster or pre-approved multi-region enablement for production workloads

Reduced re-approval friction when new model versions are introduced

This feedback is intended constructively and reflects real operational challenges faced by teams running AI Foundry as a core production platform. This issue is not a request for unlimited capacity, but for predictable, enterprise-grade operability.

Your answer

Amit Tyagi 135 Reputation points

2026-01-27T17:25:09.89+00:00

Hi Alexander,

You’re raising a very real concern that many teams hit when moving from PoC to production with Azure AI Foundry. Two key pain areas stand out:

1) TPM limits are capacity-driven, not billing-driven

Even for paid subscriptions, TPM is allocated based on regional GPU capacity and model demand, which is why limits are model and region-specific and don’t automatically carry forward to newer model versions. This works for experimentation, but becomes restrictive for mission-critical production workloads. Ref: Docs: https://learn.microsoft.com/azure/ai-services/openai/quotas-limits

2) TPM constraints block standard multi-region resiliency

Because secondary regions often lack pre-approved TPM, teams can be forced into single-region dependency, making it difficult to follow Azure’s own HA/DR best practices. During regional outages, this can lead to full service disruption even when customers are willing to design multi-region architectures. Ref: Docs: https://learn.microsoft.com/azure/architecture/framework/resiliency/overview

Bottom line: AI Foundry can run production workloads today, but only with careful capacity planning and early multi-region approvals. The current TPM and approval flow is not yet optimized for always-on, enterprise-grade resiliency — and this is exactly where platform evolution is needed.

Thanks for clearly articulating this — it’s valuable feedback for the service.
Amit Tyagi 135 Reputation points

2026-01-27T17:27:11.2933333+00:00

This issue is not a request for unlimited capacity, but for predictable, enterprise-grade operability.

Specifically, the current TPM and region-scoped approval model in Azure AI Foundry makes it difficult to design resilient, mission-critical systems, even for paying enterprise customers:

TPM approvals are model- and region-specific, requiring repeated requests with every new model version.

Limited baseline TPM prevents pre-provisioning secondary regions, blocking standard HA/DR patterns.

Regional outages (such as the recent Sweden incident) can lead to complete service disruption, even when customers are willing to operate multi-region.

This effectively places production workloads in a single-region dependency, which is inconsistent with Azure’s general reliability and resiliency guidance.

Request for consideration:

Subscription- or tenant-level TPM allocation models for enterprise customers

Faster or pre-approved multi-region enablement for production workloads

Reduced re-approval friction when new model versions are introduced

This feedback is intended constructively and reflects real operational challenges faced by teams running AI Foundry as a core production platform. This issue is not a request for unlimited capacity, but for predictable, enterprise-grade operability.